Artificial intelligence detection method for perpendicularity of working surface of spectrometer prism

Abstract

The present application relates to the technical field of optical geometry detection, and discloses an artificial intelligence detection method for perpendicularity of a working surface of a light splitting prism, comprising the following steps: processing detection physical parameters to generate a basic interference graph and build a simulation data set. An initial CNN network is built and the simulation data set is input to perform supervised learning, and an optimal target CNN network is output. In the detection link, a measurement graph is obtained from an experimental interference graph, and a pretreatment interference graph is obtained through morphological correction combined with a filtering algorithm. The pretreatment interference graph is input into the optimal target CNN network for inference, and the final fringe period is obtained by removing outliers through a threshold. According to an interference theory model, a fringe period formula and a derived prism included angle formula are derived, and the actual angle of the working surface of the light splitting prism to be measured is calculated. The present application improves the precision, anti-interference robustness and measurement stability of the light splitting prism perpendicularity detection through physical simulation and multi-region network fusion.

Description

Technical Field

[0001] This invention relates to the field of optical geometric measurement technology, specifically to an artificial intelligence method for detecting the perpendicularity of the working surface of a beam splitter prism. Background Technology

[0002] A beam splitter prism is a core optical component of a laser beam splitting and combining system. During the fabrication of a beam splitter prism, the inability to achieve absolute perpendicularity between the two working surfaces causes the transmitted and reflected beams to deviate from the design axis. Current detection methods utilize non-working wavelength laser light incident perpendicularly on the cemented surface, quantifying the perpendicularity error by capturing the equal-thickness interference fringes formed by the transmitted and reflected light.

[0003] Traditional detection methods primarily rely on manual observation of interference fringes, which is susceptible to the subjective judgment of the inspectors, resulting in low measurement accuracy and reproducibility. Although existing technologies include methods based on frequency domain analysis or the establishment of interference fringe databases for retrieval and comparison, these methods have limitations when dealing with non-ideal interference fringes. In actual detection, residual surface errors on the surface of the beam splitter under test can cause changes in the physical curvature of the fringes. Combined with thermal noise interference from photodetectors, traditional algorithms are prone to spectral leakage, leading to insufficient measurement accuracy and anti-interference capabilities.

[0004] Furthermore, existing algorithms typically perform single processing on the entire image. If there are local scratches, dust, or phase jumps caused by chamfering effects on the surface of the beam splitter, the measurement results will fluctuate, making it difficult to maintain the stability of the quantization results. At the same time, current automated inspection processes have requirements on the placement orientation of the beam splitter under test. Usually, the interference fringes need to be pre-aligned to a specific vertical or horizontal direction using a mechanical adjustment frame. Otherwise, the algorithm will have difficulty extracting geometric features. This reliance on precise manual alignment reduces the degree of automation and efficiency of the inspection.

[0005] Convolutional Neural Networks (CNNs), as a deep learning model, use multiple convolutional kernels to spatially scan images, extracting the spatial topological features of interference fringes and establishing a nonlinear mapping between image features and one-dimensional fringe periods, thus providing a means to solve the aforementioned problems. However, how to combine physical models to improve the predictive performance of neural networks in complex backgrounds remains a technical challenge in the field of optical component inspection. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an artificial intelligence detection method for the perpendicularity of the working surface of a beam splitter prism. This method solves the problems in existing technologies, such as low measurement accuracy due to weak resistance to environmental noise and surface distortion, poor measurement stability due to the influence of local surface defects, and low automation and efficiency due to reliance on manual precision alignment.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] This invention provides an artificial intelligence method for detecting the perpendicularity of the working surface of a beam splitter prism, comprising the following steps:

[0009] An optical simulation software is used to process the detected physical parameters to generate a basic interferogram. Wavefront distortion parameters are introduced into the phase field of the basic interferogram and low-order Zernike polynomial coefficients are superimposed. Gaussian noise and speckle noise are injected to obtain a perturbed interferogram. The perturbed interferogram is randomly translated horizontally and vertically to generate image samples containing slight tilt and full rotation. Physical period labels are set by combining the manual prior estimation interval to complete the construction of the simulation dataset.

[0010] An initial CNN network consisting of a feature extraction backbone and a regression prediction head is constructed. The simulation dataset is input into the network for supervised learning to establish the mapping relationship between image features and stripe periods, and the optimal target CNN network is output.

[0011] For the actual detection process, a photodetector is used to receive the interference light signal emitted from the hardware optical path and acquire an experimental interferogram. A truncated measurement map is obtained by extracting the experimental interferogram based on a multi-dimensional region evaluation function. The tilt angle of the truncated measurement map is calculated using a two-dimensional discrete Fourier transform algorithm. When the ratio of the extracted spectral main peak amplitude to the local background mean is lower than a preset signal-to-noise ratio threshold, the algorithm is switched to Radon transform to correct the tilt angle. After filtering and smoothing, a preprocessed interferogram is obtained.

[0012] Multiple local sub-images are randomly extracted from the preprocessed interferogram and input into the optimal target CNN network. The portion of the multiple period prediction values ​​that deviates from the mean by more than a preset outlier threshold is discarded as anomaly prediction extreme values. The arithmetic mean of the remaining effective prediction values ​​is calculated to obtain the final fringe period. Based on the interference theory model established on the principle of equal thickness interference, the fringe period formula and the prism angle formula derived from the fringe period formula are derived. The final fringe period, incident laser wavelength and refractive index are substituted into the formula for calculation. The constant reference value of the ideal orthogonal state is superimposed on the obtained small angle value to obtain the actual angle of the working surface of the beam splitter under test.

[0013] This invention provides an artificial intelligence method for detecting the perpendicularity of the working surface of a beam splitter prism. It has the following beneficial effects:

[0014] 1. This invention introduces Zernike polynomial distortion and noise processing into the simulation dataset, enabling the optimal target CNN network to identify distortion stripes affected by surface errors, thereby reducing the frequency leakage error of traditional frequency domain extraction algorithms in low signal-to-noise ratio environments.

[0015] 2. This invention employs a multi-region local sub-graph random sampling and outlier threshold elimination strategy, which can identify and eliminate abnormal predicted values ​​caused by edge chamfering or surface scratches. Through multi-sample fusion calculation, measurement fluctuations are smoothed out, and the repeatability of verticality error quantification is improved.

[0016] 3. This invention combines morphological expansion processing and the switching mechanism between Fourier transform and Radon transform, which can automatically respond to the randomness of the placement orientation of the beam splitter under test in the optical path, and complete the quantitative detection of perpendicularity error without manual leveling of interference fringes. Attached Figure Description

[0017] Figure 1 This is a hardware optical path structure diagram of the present invention;

[0018] Figure 2 This is a flowchart of the artificial intelligence detection method for the perpendicularity of the working surface of the beam splitter prism according to the present invention;

[0019] Figure 3 This is a flowchart illustrating the calculation of the actual angle of the working surface based on network prediction and physical model according to the present invention.

[0020] Figure 4 Distribution map of multi-region local sub-map stripe period prediction results in an application embodiment of the present invention.

[0021] Among them, 1. laser source; 2. first working surface; 3. second working surface; 4. exit surface; 5. incident surface; 6. adhesive surface; 7. photodetector; 8. beam splitter to be tested. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see the appendix Figure 1 and attached Figure 2 This invention provides an artificial intelligence method for detecting the perpendicularity of the working surface of a beam splitter prism. This method is based on a hardware optical path for data acquisition. The hardware optical path mainly consists of a laser source 1, a beam splitter prism 8 under test, and a photodetector 7. The laser source 1 provides a non-working band laser of a specific wavelength, which is then perpendicularly incident on the cemented surface 6 of the beam splitter prism 8 under test via the incident surface 5.

[0024] Furthermore, a beam-splitting film is coated on the adhesive surface 6, which splits the incident non-working wavelength laser beam into a transmitted beam and a reflected beam. Due to the non-perpendicularity of the first working surface 2 and the second working surface 3 during the fabrication of the beam-splitting prism 8, the transmitted and reflected beams reach the first working surface 2 and the second working surface 3 respectively and undergo residual reflection. The two residual reflected beams overlap in space, producing interference fringes of equal thickness. A photodetector 7 is placed in the optical path of the exit surface 4 to receive the interfered optical signal and convert it into an electrical signal.

[0025] The method includes the following steps:

[0026] Step S1: Construct and enhance the simulation interferometric dataset. Externally input detection physical parameters are processed using optical simulation software to generate a basic interferogram. Subsequently, multiple physical perturbations are applied to the basic interferogram to enhance the data, obtaining a perturbed interferogram. Based on this perturbed interferogram, its shape is further expanded and target labels are set to complete the construction of the simulation dataset.

[0027] After completing the dataset construction, proceed to step S2 to build and train the convolutional neural network. An initial CNN network (convolutional neural network) is constructed, and the aforementioned simulation dataset is input into the initial CNN network for supervised learning. The initial CNN network extracts the spatial topological features of the interference fringes and iteratively updates the weight parameters, establishing a mapping relationship between image features and fringe periods, and finally outputs the optimal target CNN network.

[0028] For the actual testing process, step S3 is executed to acquire and preprocess the experimental interferometric image. The photodetector 7 receives the interference light signal emitted from the hardware optical path and converts it into an electrical signal, thus acquiring the experimental interferogram. To eliminate interference from local surface errors, a region with uniform fringe distribution is selected from the experimental interferogram to obtain a truncated measurement image. Next, morphological correction is performed on the truncated measurement image, and smoothing and denoising are carried out using a filtering algorithm to obtain a preprocessed interferogram.

[0029] Based on the above preprocessing results, step S4 is executed to calculate the actual angle based on network prediction and the physical model. The preprocessed interferogram is input into the optimal target CNN network to predict the final fringe period. Based on this, the interference parameter correlation is constructed according to the externally input interference theory model, where the optical path difference generated by the remaining reflected light at the meeting point satisfies the optical path difference formula:

[0030] ;

[0031] In the formula: The optical path difference between the two remaining reflected beams at the point of intersection; The refractive index of the beam splitter 8 to be tested is given. This represents the thickness of the medium at the corresponding location.

[0032] The spatial distribution of interference light intensity follows the interference light intensity formula:

[0033] ;

[0034] In the formula: The intensity of the interference light received on the target surface of photodetector 7; The intensity of the remaining reflected light from the first beam that participates in the interference; The intensity of the remaining reflected light from the second beam that participates in the interference; Pi is a constant. The incident laser wavelength; This represents the optical path difference generated at the point where the two remaining reflected beams meet.

[0035] The physical medium thickness condition corresponding to the formation of bright fringe centers during constructive interference conforms to the bright fringe center formula:

[0036] ;

[0037] In the formula: The refractive index of the beam splitter 8 to be tested is given. This represents the thickness of the medium at the corresponding location; Let be the interference series, and its value be a non-negative integer. The incident laser wavelength is denoted as .

[0038] By deriving the evolution of thickness difference, a mathematical relationship is established between the final fringe period and the minute included angle, and the formula for the fringe period is derived:

[0039] ;

[0040] In the formula: The refractive index of the beam splitter 8 to be tested is given. The period of the interference fringes; The small angle between the two working surfaces of the beam splitter 8 to be tested; The incident laser wavelength is denoted as .

[0041] After obtaining the final fringe period, substitute it into the prism angle formula:

[0042] ;

[0043] In the formula: The actual angles between the two working surfaces of the beam splitter 8 to be tested; The incident laser wavelength; The refractive index of the beam splitter 8 to be tested is given. The period of the interference fringes is denoted by .

[0044] The actual angle of the working surface is calculated using the arctangent function, thus completing the quantitative detection of the perpendicularity error of the working surface of the beam splitter 8 under test.

[0045] In this embodiment, the specific process of constructing and enhancing the simulation interferometric dataset in step S1 is discussed in detail.

[0046] As the data basis of this method, in step S101, optical simulation software is used to process the externally input detection physical parameters and generate a basic interferogram accordingly.

[0047] Specifically, the optical simulation software used here can preferably be a numerical computing platform with physical optical tracing capabilities (such as MATLAB software, Zemax optical design software, etc.). The physical parameters to be detected specifically cover the specific detection wavelength used by the beam splitter under test in the actual detection environment, as well as the refractive index of the beam splitter material at the corresponding wavelength.

[0048] The operator inputs the refractive index, wavelength, and initial interference angle into the simulation model, and the simulation software can calculate the interference intensity distribution matrix in two-dimensional space based on the theoretical light intensity distribution law of equal thickness interference.

[0049] To adapt to the fixed-dimensional input interface of subsequent convolutional neural networks, the light intensity distribution matrix needs to be mapped to 8-bit or 16-bit grayscale image data at a specific resolution (e.g., 256×256 pixels), thus forming a visualized two-dimensional basic interferogram. This basic interferogram represents the absolutely straight interference fringes on an ideal smooth optical surface and in a clean environment, serving as the data basis for subsequent perturbation injection. For the specific numerical solution algorithm for calculating the interference light intensity distribution matrix in optical simulation software, those skilled in the art can consult relevant optical computation literature for implementation; the numerical solution of the light field distribution is a well-known technique in this field and will not be elaborated upon here.

[0050] Considering the feature domain deviation between the purely theoretical simulation image and the physical image acquired by the actual photodetector 7, step S102 is executed to perform multiple physical perturbation processing on the basic interferogram for data enhancement, thereby obtaining a perturbation interferogram.

[0051] In practice, wavefront distortion parameters are introduced to simulate the fringe curvature changes caused by the uneven surface profile of the beam splitter under test. As a preferred distortion simulation method, these wavefront distortion parameters are specifically constructed by introducing Zernike polynomials. By superimposing different low-order Zernike polynomial coefficients (e.g., the first 4th to 9th order polynomials representing defocus, astigmatism, coma, and spherical aberration) into the simulated phase field, the residual surface profile errors of the beam splitter surface in the actual optical processing can be effectively simulated, causing the original ideal straight interference fringes to produce edge curvature and local distortion that conform to physical laws.

[0052] Based on this distorted image, Gaussian noise, speckle noise, and uneven background illumination are randomly injected into the interferogram.

[0053] From the perspective of physical causes and mathematical models, speckle noise is used to reproduce the granular random interference spots generated when highly coherent lasers are scattered on rough optical surfaces. It is usually superimposed on the original image matrix using a multiplicative noise model. Gaussian noise is used to simulate the unavoidable thermal noise and dark current noise of a real photodetector 7 when performing photoelectric signal conversion. Non-uniform background illumination is used to simulate the characteristics of the energy distribution of the Gaussian beam emitted by a real laser, which is strong at the center and weak at the edges, as well as stray light from the external environment. Local brightness modulation can be achieved by superimposing a two-dimensional Gaussian distribution function.

[0054] The intensity coefficients for the aforementioned noise and background injection are set based on the actual hardware signal-to-noise ratio parameters of the photodetector 7 and the ambient illumination of the test environment. For example, the variance range of the Gaussian noise is set to a random value between 0.01 and 0.05 to approximate the background noise level of an industrial-grade CMOS camera. After the above-mentioned multiple physical perturbation processes, a perturbation interferogram with visual features highly similar to the image acquired in the actual physical experiment can be output.

[0055] After obtaining the perturbation interferogram, step S103 is executed to perform morphological augmentation on the perturbation interferogram and assign target labels to the augmented image to complete the construction of the simulation dataset. Given the randomness of the position and orientation of the beam splitter under test in the optical path during actual detection, the algorithm performs random horizontal and vertical translation operations on images of the same period. The translation range can be set to 0% to 10% of the image width and height. This operation aims to give the subsequent network model translation invariance to the target position.

[0056] To address the angular characteristics of the fringe space, the algorithm generates image samples that include both slight tilt and full rotation. In this process, when rotating the discrete image matrix, to avoid jagged edges caused by grid point misalignment, bilinear or bicubic interpolation algorithms are used to resample pixels. The preset range for slight tilt angles is [-2°, 2°]. This preset range is based on the precise alignment tolerance of the actual optical adjustment frame, for example, the residual artificial tilt angle when the inspector uses a fine-tuning knob to coarsely adjust the interference fringes to a roughly horizontal or vertical state. The preset range for full rotation angles is [-90°, 90°], designed to cover all possible interference fringe orientations without any artificial angle coarse adjustments.

[0057] After generating a massive number of image samples through the aforementioned morphological augmentation process, the target labels for these images are uniformly set to the physical period floating within a manually estimated range. This manually estimated range is determined by inversely calculating the perpendicularity tolerance range marked on the beam splitter's manufacturing drawings. For example, if the design perpendicularity tolerance of the beam splitter under test is ±5 arcminutes, and considering the detection wavelength, the fringe period range presented on the detector target surface is approximately 10 to 150 pixels, calculated using interference theory. Images with corresponding period labels are generated within this range. This completes the construction of a simulation dataset encompassing multi-dimensional physical features and bearing one-dimensional period numerical labels. This simulation dataset is stored as standard data input for subsequent network supervised learning.

[0058] In this embodiment, based on the simulation dataset that has been constructed as described above, the core construction and parameter optimization of the detection model are further performed, namely, step S2, which involves constructing and training a convolutional neural network.

[0059] Considering the highly nonlinear nature of the mapping from the interference fringe image to the geometric parameters, an initial CNN network for the regression prediction task is constructed during step S201.

[0060] Specifically, in order to establish an effective mapping from two-dimensional physical images to one-dimensional geometric constants, the input of the initial CNN network is set to a two-dimensional image data interface adapted to the image dimension of the simulation dataset (e.g., receiving a grayscale image tensor with a resolution of 256×256 pixels, a single-channel grayscale value of 0 to 255, and normalized by dividing the pixel value by 255), and the output is set to a fully connected regression node that outputs one-dimensional numerical prediction results.

[0061] As a preferred network architecture, this initial CNN network consists of a feature extraction backbone and a regression prediction head. The feature extraction backbone uses alternating convolutional blocks, with each convolutional block arranged in the order of "convolutional layer, batch normalization layer, activation function layer".

[0062] The convolutional layers are configured with multiple kernels of specific sizes (such as 3×3 or 5×5 pixels). They extract the spatial topological features of interference fringes by performing sliding window-like dot product operations on the input image. Based on optical imaging features, shallow convolutional kernels are mainly used to capture edge gradients and local gray-level abrupt changes in the fringes, while deep convolutional kernels are used to integrate the global fringe density and overall tilt. Batch normalization layers normalize the mean and variance of each feature map, aiming to accelerate network convergence and mitigate internal covariate shifts. The activation function layers preferably use the Rectified Linear Unit (ReLU) function to introduce non-linear expressive power into the network. Pooling layers progressively reduce the spatial resolution of the feature maps through downsampling operations, reducing the number of network parameters while preserving core features.

[0063] Global average pooling layers and fully connected layers are connected at the end of the backbone network to flatten the high-dimensional feature map and map it to the stripe periodicity numerical space. For the specific feature map dimension calculation and tensor transfer mechanism of the bottom layer of the convolutional neural network, those skilled in the art can refer to relevant literature on deep learning. The network node operation principle is a well-known technology in this field and will not be elaborated here.

[0064] After the basic architecture is built, proceed to step S202, where the simulation dataset constructed in step S103 is input into the initial CNN network for supervised learning.

[0065] During the training phase, the initial CNN network uses the forward propagation algorithm to process the input simulation dataset images and calculates the current predicted stripe period. This predicted stripe period is then compared with the physical period label provided in the simulation dataset, and the error between the two is calculated using a preset loss function. Given that this scheme is a continuous numerical regression prediction task, the mean squared error loss function or the smoothed L1 loss function is preferred.

[0066] In the specific calculation, the predicted period value is subtracted from the label period value and the result is squared or calculated as an absolute value. The resulting value directly represents the degree of deviation between the model prediction result and the actual physical distance.

[0067] To achieve adaptive optimization of network parameters, a backpropagation algorithm combined with a gradient descent optimizer (such as the Adam optimizer) is employed. During optimizer calculations, to prevent computational crashes caused by the denominator approaching zero, a very small constant correction value (e.g., 10 to the power of -8) is introduced into the denominator. Based on the acquired gradient information, the network parameters are iteratively updated according to a preset learning rate.

[0068] The preset learning rate is set based on the trade-off between the convergence speed in the early stage of network training and the optimization stability in the later stage of training. For example, a dynamic decay learning rate strategy can be adopted, with the initial learning rate set at 0.001 to accelerate the escape from local optima, and gradually decayed to 0.0001 after several training rounds to achieve fine-grained convergence.

[0069] Meanwhile, the training process adopts a mini-batch input strategy. The preset batch size is set according to the device's video memory capacity and the smoothness of the gradient descent direction, for example, it is configured as 32 or 64 images per batch.

[0070] As the number of training rounds increases, the initial CNN network establishes a robust nonlinear mapping relationship from the features of two-dimensional interference images with noise and distortion to the period of one-dimensional interference fringes through continuous stimulation by a massive number of samples with perturbation features.

[0071] The network's loss value on the independent validation set is monitored in real time. When the validation set loss value no longer decreases within a preset tolerance number of epochs (e.g., 10 consecutive epochs), or reaches the preset maximum training epoch limit, the network is considered to have reached convergence. Training is terminated, and the network weight parameters at this point are fixed. Finally, the optimal target CNN network is output. This optimal target CNN network will serve as the core inference tool for subsequent accurate periodic predictions based on real physical experimental data.

[0072] In this embodiment, the actual physical measurement and data processing stage is entered, namely, step S3, which involves acquiring and preprocessing experimental interference images.

[0073] In step S301, an external photodetector 7 receives the interference light signal emitted from the hardware optical path and converts it into an electrical signal, thereby acquiring the experimental interferogram. Specifically, the photodetector 7 preferably uses an industrial-grade CCD camera or a CMOS image sensor, with its photosensitive target surface placed perpendicularly on the output optical path of the beam splitter 8 under test. The photodetector 7 converts the spatial distribution characteristics of the received interference light intensity into an analog electrical signal through its internal photoelectric effect, which is then processed by an analog-to-digital converter unit, ultimately outputting digitized two-dimensional global image data, i.e., the experimental interferogram.

[0074] Considering the edge chamfering effect of the beam splitter 8 under test during actual fabrication, and the physical characteristics of the incident Gaussian beam itself—high energy at the center and weak energy at the edges—the edge regions of the global experimental interferogram often exhibit significant local surface shape errors and light intensity attenuation. To eliminate these physical interferences and match the input interface of the network model, the edge distortion regions are removed from the experimental interferogram, and an effective interferometric region with a size larger than the network input dimension is selected to obtain the truncated measurement map.

[0075] In practical engineering operations, considering that relying on a single local contrast extremum for cropping is easily affected by dust on the optical mirror or scratches on the local surface, a multi-dimensional region evaluation function is established to determine the optimal cropping coordinates. By traversing the global experimental interferogram through a sliding window, the gray-level variance of pixels within each window is extracted as a contrast feature, and the Euclidean distance from the center of the current window to the geometric center of the global effective aperture is calculated as a positional deviation feature.

[0076] Furthermore, a spatial penalty coefficient is set for the distance feature, and it is then weighted and summed with the contrast feature. When normalizing the weighting factors, a minimum constant (e.g., 10 to the power of -8) is added to the denominator to prevent the denominator from approaching zero and causing computational overflow. Finally, based on the maximum coordinate of this multi-dimensional weighted evaluation function, the region that balances high stripe contrast with minimal physical deviation is selected as the intercepted measurement map.

[0077] After acquiring the intercepted measurement image, to eliminate the randomness of the placement orientation of the photodetector 7 or the beam splitter 8 under test in the experimental environment, step S302 is executed to perform morphological correction on the intercepted measurement image. To align with the training benchmark of the optimal target CNN network constructed in the early stage and to reduce the spatial search complexity of the model during the inference stage, the overall orientation of the interference fringes in the image is corrected to a vertical or horizontal state through a rotation operation.

[0078] As a preferred implementation, a two-dimensional discrete Fourier transform algorithm is used to transform the intercepted measurement image from the spatial domain to the frequency domain. By locating the coordinates of the dominant frequency pulse with the most concentrated energy in the spectrum image, the initial tilt angle of the current interference fringes is calculated. During this angle calculation process, if the ratio of the extracted spectral peak amplitude to the local background mean is lower than a preset signal-to-noise ratio threshold, it indicates that the current frequency domain features are singular or masked by strong noise. In this case, the algorithm is adaptively switched to the Radon transform algorithm, which projects and integrates the image along multiple discrete angles, and selects the angle with the largest integral variance as the final tilt angle.

[0079] Based on the calculated tilt angle, an affine rotation transformation is performed on the discrete image matrix using a bilinear interpolation algorithm to eliminate the original tilt deviation. Simultaneously, to address the physical phenomenon of the image corners extending beyond the original matrix boundaries caused by the affine rotation, an edge pixel mirroring extension strategy is employed for boundary completion. This ensures that the tensor size of the resampled image does not shrink and does not introduce abrupt high-frequency artifacts.

[0080] After the aforementioned morphological correction and resampling interpolation, a small amount of interpolation artifacts will inevitably be introduced into the image, and the cropped measurement image itself also carries high-frequency hardware thermal noise generated in the previous photoelectric conversion process. Based on these considerations, a filtering algorithm is used to smooth and denoise the corrected image, thereby suppressing high-frequency detection noise and improving the contrast between adjacent bright and dark fringes, resulting in a preprocessed interferogram.

[0081] In practical applications, Gaussian smoothing or median filtering algorithms can be used. Implementing this filtering algorithm involves setting a preset filtering window size, which is determined based on the ratio between the physical pixel size of the detector and the physical width of the interference fringes on the target surface. For example, when a single bright fringe spans approximately 20 pixels, the preset filtering window size can be set to a 3×3 or 5×5 pixel array. This setting logic aims to ensure effective mean smoothing of random background noise without blurring or erasing the edge features of the physical interference fringes due to an excessively large window.

[0082] After going through the above acquisition, cropping, correction and denoising process, a standardized preprocessed interferogram is finally output. This image will serve as the sole input data for network inference and prediction in subsequent steps.

[0083] See appendix Figure 3 In this embodiment, after obtaining the preprocessed interferogram, the final calculation stage of transforming visual feature extraction into physical geometric parameters is entered, namely, step S4, which calculates the actual included angle based on network prediction and physical model.

[0084] In this embodiment, to obtain high-confidence fringe geometric parameters, step S401 is executed, in which the obtained preprocessed interferogram is input into the trained optimal target CNN network for inference prediction. To further improve prediction accuracy and robustness in complex detection environments, a multi-region sampling fusion strategy is adopted.

[0085] In practice, within the effective interference region of the same global experimental interferogram, multiple local sub-images at different locations are cropped according to random coordinate starting points. The cropping size of the local sub-images is adapted to the fixed input dimension of the optimal target CNN network. These local sub-images are then input into the optimal target CNN network, and the network outputs multiple discrete periodic prediction values ​​through forward propagation.

[0086] To account for the possibility of outliers arising from drastic distortions during the measurement process, an outlier removal mechanism is introduced before calculating the average. The preset outlier threshold is set based on the confidence interval of the normal distribution. For example, the mean and standard deviation of the predicted values ​​are calculated, and outlier predicted extreme values ​​that deviate from the mean by more than 3 times the standard deviation are removed.

[0087] In this statistical calculation, if the standard deviation is detected to be close to zero (e.g., the value is less than 10 to the power of -5), it indicates that the prediction results of each local subgraph are highly consistent, and the anomaly removal logic will be skipped directly to avoid the potential risk of dividing the value by zero; at the same time, the number of remaining valid prediction values ​​after removal is checked. If the number is lower than the preset minimum number of valid samples (e.g., 3), a resampling instruction is triggered.

[0088] After confirming the reliability of the sample distribution, the arithmetic mean of the remaining valid predicted values ​​after removing outliers is calculated to obtain the final fringe period. This multi-region sampling fusion strategy effectively mitigates the interference of possible residual local phase distortion in a single region on the overall period prediction.

[0089] After obtaining the final fringe period, step S402 is executed to construct the interference parameter correlation based on the externally input physical theoretical model. This physical theoretical model is based on the principle of equal-thickness interference. When two beams of residual reflected light meet in the probe space, the resulting optical path difference determines the spatial distribution of the bright and dark areas of the interference fringes. This optical path difference satisfies the optical path difference formula:

[0090] ;

[0091] In the formula: The optical path difference between the two remaining reflected beams at the point of intersection; The refractive index of the beam splitter 8 to be tested is given. This represents the thickness of the medium at the corresponding location.

[0092] After the two coherent beams are superimposed, the spatial distribution of light intensity received on the target surface of photodetector 7 follows the interference light intensity formula:

[0093] ;

[0094] In the formula: The intensity of the interference light received on the target surface of photodetector 7; The intensity of the remaining reflected light from the first beam that participates in the interference; The intensity of the remaining reflected light from the second beam that participates in the interference; Pi is a constant. The incident laser wavelength; This represents the optical path difference generated at the point where the two remaining reflected beams meet.

[0095] In the interference pattern, when the cosine term in the light intensity formula reaches its maximum value (i.e., the phase difference is...), When the light beams are multiples of each other (integer multiples of each other), they interfere constructively, forming bright interference fringe centers. The physical medium thickness condition corresponding to these centers conforms to the formula for bright fringe centers:

[0096] ;

[0097] In the formula: The refractive index of the beam splitter 8 to be tested is given. This represents the thickness of the medium at the corresponding location; Let be the interference series, and its value be a non-negative integer. The incident laser wavelength is denoted as .

[0098] Based on the above conditions for bright fringe centers, by analyzing the evolution of the medium thickness difference corresponding to the centers of adjacent bright fringe levels, and combining this with geometric trigonometric functions for derivation, a periodicity is established. With a tiny angle The mathematical correlation between optical coherence and device physical structure characteristics leads to the formula for the fringe period:

[0099] ;

[0100] In the formula: The refractive index of the beam splitter 8 to be tested is given. The period of the interference fringes; The small angle between the two working surfaces of the beam splitter 8 to be tested; The incident laser wavelength is denoted as .

[0101] Based on the mathematical correlation derived above, step S403 is executed to obtain the final fringe period, which is then substituted into the prism angle formula derived from the fringe period formula:

[0102] ;

[0103] In the formula: The actual angles between the two working surfaces of the beam splitter 8 to be tested; The incident laser wavelength; The refractive index of the beam splitter 8 to be tested is given. The period of the interference fringes is denoted by .

[0104] In this calculation step, the final fringe period, the incident laser wavelength and refractive index corresponding to the external input detection environment are substituted into the prism angle formula.

[0105] Considering the risk of underflow when computers perform floating-point division operations, and the fact that the refractive index of real physical optical media is always greater than 1, a non-zero division check logic is added to the operation logic to determine whether the final fringe period approaches zero.

[0106] When the predicted interference fringe period in the denominator term abnormally approaches zero (e.g., due to network output failure caused by extreme background light interference), an out-of-bounds warning is thrown and the current solution process is interrupted to ensure the completeness and robustness of the algorithm.

[0107] After verification, the arctangent function is calculated on the parameter combination consisting of the final fringe period, incident laser wavelength, and refractive index to obtain the value of the minute included angle. After obtaining the value of the minute included angle, it is equivalent to the perpendicularity error amplitude. Then, a 90-degree constant reference value of the ideal orthogonal state is superimposed on the value of the minute included angle to obtain the actual angle of the working surface. This completes the quantitative detection of the perpendicularity error of the working surface of the beam splitter 8 under test.

[0108] To further aid in understanding the technical solution of the present invention, the following detailed description is provided in conjunction with a specific industrial testing application embodiment.

[0109] In this embodiment, the object to be tested is an N-BK7 material beam splitter prism 8 with a design standard of 90 degrees orthogonal. The laser source 1 configured for the testing environment emits wavelengths... The wavelength is 632.8 nm (0.6328 μm), which is outside the operating band for the beam splitter in this embodiment. The refractive index of N-BK7 material at this wavelength is known. The value is 1.515. The photodetector 7 uses a pixel physical size... It is an industrial-grade CMOS camera with a 3.45μm aperture.

[0110] In the data acquisition and preprocessing stage, the laser source 1 is first controlled to emit a beam, which, after hardware optical path interference, is acquired by the photodetector 7 with a resolution of [missing information]. Global experimental interferogram of pixels. Considering the distortion at the edges of the field of view, edge regions are removed, and the retained size is [value missing]. The effective interference region of a pixel is used as the intercepted measurement map.

[0111] The tilt angle was calculated by performing a two-dimensional discrete Fourier transform on the intercepted measurement image, and rotation correction and edge mirroring were performed using bilinear interpolation. Then, an application was made... Median filtering of the window is used for smoothing and denoising to finally obtain a standard preprocessed interferogram.

[0112] After obtaining the preprocessed interferogram, the system proceeds to the network prediction and physical model calculation stage. The system randomly selects five dimensions from the preprocessed interferogram, based on random coordinates, to fit the optimal target CNN network's fixed dimensions (e.g., ...). A local sub-image (pixels).

[0113] These five local sub-images (local sub-images numbered 1-5) were input into the trained optimal target CNN network. The predicted interference fringe period values ​​output by the network were 45.2 pixels, 45.1 pixels, 44.9 pixels, 45.3 pixels, and 45.0 pixels, respectively. The mean and standard deviation of this set of predicted values ​​were calculated to confirm that none exceeded the preset outlier threshold. Then, the arithmetic mean of these five valid predicted values ​​was calculated, yielding the final fringe period pixel value of 45.1 pixels. Combining this with the physical size of the photodetector 7 pixels, the interference fringe period was converted into a physical dimension. .

[0114] Based on the above physical parameters, the system calculates the actual angle of the working surface according to the derived formula for the prism angle. The calculation process by substituting the data is as follows:

[0115] ;

[0116] ;

[0117] ;

[0118] ;

[0119] In this calculation step, the dimensionless ratio of 0.001342 within the parentheses of the arctangent function is the tangent of the minute angle between the working surfaces. The calculated 0.0769° (approximately 4.61 arcminutes) is the perpendicularity error amplitude. Combining this with the priori rules for determining the relative movement direction of the interference fringes, this error amplitude is added to the 90° constant reference value to finally quantify the actual angle of the working surface of the beam splitter 8 under test. It is 90.0769°.

[0120] See appendix Figure 4 , Figure 4 The horizontal axis is set as the local sub-image number, with a scale range of discrete integers from 1 to 5, corresponding to 5 randomly selected local sub-images input into the optimal target CNN network; the vertical axis is set as the predicted stripe period, in pixels, with a scale range covering 44.8 to 45.4 pixels.

[0121] Figure 4 It contains five independent black square scatter plots, corresponding to coordinates (1, 45.2), (2, 45.1), (3, 44.9), (4, 45.3), and (5, 45.0), reflecting the fluctuations in predicted values ​​caused by minor differences in the physical surface features of each local subplot. Meanwhile, Figure 4 A horizontal black dashed line with a vertical coordinate of 45.1 runs through the center. This dashed line represents the final fringe period baseline calculated by taking the arithmetic mean of the five valid prediction values. By comparing the scatter points with the horizontal dashed line, it is shown how this scheme uses a multi-region fusion algorithm to smooth out single prediction fluctuations caused by local distortions and output high-confidence physical period parameters.

[0122] To further verify the effectiveness and advantages of the technical solution provided by this invention, experimental verification and effect comparison were carried out.

[0123] Experimental conditions and sample setup:

[0124] Fifty N-BK7 beam splitters manufactured in the same batch were selected as experimental samples. The measurement results, obtained using a high-precision photoelectric autocollimator with a nominal measurement accuracy of 0.5 arcseconds combined with a precision turntable, were used as the physical reference true value for the actual included angle of the working surface perpendicularity of the experimental samples. To simulate the imperfect conditions in the industrial testing environment, random environmental stray light interference with a variance of 0.02 to 0.05 was set in the hardware optical path, and trace amounts of processing dust residue were retained at the optical surface edges of some of the beam splitters under test to reproduce local surface distortion and high-frequency noise.

[0125] Comparison method design:

[0126] The above 50 experimental samples were placed under the same hardware optical path to acquire experimental interferograms, and three different algorithms were used for processing and comparison:

[0127] Control group 1 (traditional frequency domain algorithm): The two-dimensional fast Fourier transform (FFT) algorithm is used to directly extract the spectral main peak coordinates of the global experimental interferogram to calculate the fringe period, and then substitute it into the physical formula for solution.

[0128] Control group 2 (conventional full-field CNN algorithm): A single full-field image is directly input into a conventional convolutional neural network for prediction. The network training set is a pure, ideal straight-line interferogram, and no multi-region sampling and outlier removal strategies are applied during the inference stage.

[0129] Experimental group (method of the present invention): adopts the complete steps of the present invention, including training the optimal target CNN network based on a simulation dataset enhanced by multiple physical perturbations, and multi-region sampling fusion and outlier removal strategies in the inference stage.

[0130] Experimental data statistics and calculation steps:

[0131] For 50 experimental samples, the true physical reference value measured by the photoelectric autocollimator and the actual working surface angle output by three detection methods were obtained. The absolute value of the difference between the actual working surface angle and the true physical reference value was uniformly converted to the order of arcseconds and denoted as the angle error of a single sample. The specific calculation steps for the experimental data indicators are as follows:

[0132] Calculation of mean absolute error: The angle errors generated by a certain detection method on all samples that successfully output measurement results are summed, and then divided by the total number of samples that successfully output measurement results to obtain the arithmetic mean, which is used as the mean absolute error.

[0133] Calculation of standard deviation of error: Calculate the difference between the angle error and the mean absolute error for each successful measurement sample, sum the squares of the difference, divide by the total number of samples with successful output measurement results, and finally take the square root of the quotient to obtain the standard deviation of error.

[0134] Calculation of effective detection rate under complex interference: A pre-set industrial tolerance threshold is used (e.g., an angle error threshold of 30 arcseconds). The total number of valid samples for a given detection method out of 50 experimental samples is calculated, provided the calculation program does not crash due to interference and the angle error of its output does not exceed the angle error threshold. This valid sample count is divided by the total number of experimental samples (50), and the ratio is converted to a percentage. This percentage is then used as the effective detection rate under complex interference.

[0135] The comparison results obtained based on the above steps are shown in Table 1.

[0136] Table 1. Comparison of experimental data from different detection methods

[0137] Detection methods Mean absolute error (arcseconds) Standard deviation of error (arcseconds) Effective detection rate under complex interference Control Group 1 (Traditional Frequency Domain Algorithm) 12.4 5.8 68.00% Control group 2 (conventional full-field CNN algorithm) 6.5 3.2 82.00% Experimental group (method of this invention) 1.8 0.7 98.00%

[0138] According to the statistical data in Table 1, the technical advantages of this invention include:

[0139] The mean absolute error of control group 1 reached 12.4 arcseconds, indicating that traditional frequency domain algorithms are prone to spectral leakage and period extraction failure when faced with fringe distortion caused by environmental stray light and surface dust. The mean absolute error of the experimental group decreased to 1.8 arcseconds, verifying the effectiveness of the present invention in introducing Zernike polynomial wavefront distortion parameters and multiple physical noises for training in the simulation dataset. This operation enables the optimal target CNN network to acquire feature perception capabilities against optical processing errors and thermal noise, and can identify fringe topology structures in high-noise backgrounds.

[0140] Comparing the data from control group 2 and the experimental group, the standard deviation of the error decreased from 3.2 arcseconds to 0.7 arcseconds. Because control group 2 uses a single global prediction, the phase jump directly affects the output value of the global period when local dust or edge chamfering effects exist in the interferogram. This invention employs a multi-region local sub-graph random sampling and outlier threshold removal strategy. Through multi-sample fusion calculation, it eliminates abnormal prediction extreme values ​​caused by local defects, smooths out fluctuations in single predictions, and ensures the repeatability of the quantization results.

[0141] In a test environment with introduced interference, the experimental group achieved an effective detection rate of 98.0%, outperforming traditional detection methods. By incorporating an adaptive angle correction mechanism in morphological correction, the system of this invention can automatically adapt to the randomness of the placement orientation of the beam splitter under test, reducing reliance on precise manual alignment and providing a foundation for automated inspection in actual industrial manufacturing production lines.

Claims

1. An artificial intelligence method for detecting the perpendicularity of the working surface of a beam splitter prism, characterized in that, Includes the following steps: Optical simulation software is used to process and detect physical parameters to generate a basic interferogram. Multiple physical perturbations are applied to the basic interferogram to obtain a perturbed interferogram. The perturbed interferogram is then morphologically expanded and target labels are set to construct a simulation dataset. An initial CNN network is constructed, and the simulation dataset is input into the initial CNN network for supervised learning. The mapping relationship between image features and stripe periods is established, and the optimal target CNN network is output. The experimental interferogram is obtained by receiving the interference light signal using the photodetector (7) in the hardware optical path, extracting the truncated measurement image from the experimental interferogram, performing morphological correction on the truncated measurement image and processing it using a filtering algorithm to obtain the preprocessed interferogram; Multiple local sub-images are extracted from the preprocessed interferogram, and the local sub-images are respectively input into the optimal target CNN network for prediction. After removing outliers, the arithmetic mean is taken to calculate the final stripe period. Based on the interference theory model, the interference parameter correlation is constructed to derive the fringe period formula. The final fringe period is substituted into the prism angle formula derived from the fringe period formula, and the actual angle of the working surface is calculated by the arctangent function to complete the detection. The steps of performing morphological correction on the intercepted measurement image and processing it using a filtering algorithm to obtain a preprocessed interferogram specifically include: The intercepted measurement image is converted to the frequency domain using a two-dimensional discrete Fourier transform algorithm, and the tilt angle is calculated by locating the coordinates of the main frequency pulse. When the ratio of the amplitude of the main peak of the spectrum extracted from the frequency domain to the mean of the local background is lower than the preset signal-to-noise ratio threshold, the algorithm is switched to the Radon transform algorithm to project and integrate along multiple discrete angles, and the angle with the largest integral variance is selected as the final tilt angle. Based on the final tilt angle calculated, an affine rotation transformation is performed on the discrete image matrix corresponding to the intercepted measurement map using a bilinear interpolation algorithm, and an edge pixel mirror extension strategy is used to complete the boundary of the intercepted measurement map. The filtering algorithm is used to smooth and denoise the corrected and completed truncated measurement image to obtain the preprocessed interferogram; The preset signal-to-noise ratio threshold is set based on the relative distribution characteristics of the amplitude of the main peak of the spectrum and the mean of the local background. The steps of extracting multiple local sub-images from the preprocessed interferogram, inputting each local sub-image into the optimal target CNN network for prediction, and calculating the final fringe period by taking the arithmetic mean after removing outliers specifically include: Within the preprocessed interferogram, multiple local sub-images at different positions are cropped according to random coordinate starting points; The multiple local sub-graphs are respectively input into the optimal target CNN network, and multiple discrete periodic prediction values ​​are output; Calculate the mean and standard deviation of the periodic predicted values, and remove the portion of the periodic predicted values ​​that deviates from the mean by more than a preset outlier threshold as abnormal predicted extreme values. The remaining periodic prediction values ​​after removing the abnormal prediction extreme values ​​are taken as effective prediction values. The arithmetic mean of the effective prediction values ​​is calculated to obtain the final stripe period. The preset outlier threshold is set based on the normal distribution confidence interval.

2. The artificial intelligence detection method for the perpendicularity of the working surface of a beam splitter prism according to claim 1, characterized in that, The hardware optical path also includes a laser source (1) and a beam splitter (8) to be tested. The laser source (1) is used to provide non-working band laser, so that the non-working band laser is incident perpendicularly onto the cemented surface (6) of the beam splitter (8) under test through the incident surface (5). A beam splitting film is coated on the adhesive surface (6). The beam splitting film is used to split the incident non-working band laser into a transmitted beam and a reflected beam. The transmitted beam and the reflected beam reach the first working surface (2) and the second working surface (3) respectively and undergo residual reflection. The two residual reflected beams overlap in space to generate equal-thickness interference fringes to form the interference light signal. The photodetector (7) is placed in the optical path of the exit surface (4) to receive the interference light signal and convert it into an electrical signal.

3. The artificial intelligence detection method for the perpendicularity of the working surface of a beam splitter prism according to claim 1, characterized in that, The steps of performing multiple physical perturbation processing on the basic interferogram to obtain a perturbed interferogram specifically include: Wavefront distortion parameters are introduced into the phase field of the basic interferogram and low-order Zernike polynomial coefficients are superimposed to simulate the fringe curvature change caused by the uneven surface profile of the beam splitter (8) under test. In the basic interferogram after introducing the wavefront distortion parameters, the perturbation interferogram is obtained by randomly injecting Gaussian noise, speckle noise and non-uniform background illumination using a multiplicative noise model and a two-dimensional Gaussian distribution function.

4. The artificial intelligence detection method for the perpendicularity of the working surface of a beam splitter prism according to claim 1, characterized in that, The specific steps for morphological augmentation and target labeling of the perturbation interferogram to construct the simulation dataset include: Perform random horizontal and vertical translation operations on the perturbation interferogram; Generate image samples containing tilt and full rotation, and uniformly set the target label of the expanded perturbation interferogram to the physical period of the floating within the artificial prior estimation interval, thus completing the construction of the simulation dataset; The artificial prior estimation range is set by back-calculation based on the perpendicularity tolerance range marked on the processing drawing of the beam splitter (8) to be tested.

5. The artificial intelligence detection method for the perpendicularity of the working surface of a beam splitter prism according to claim 1, characterized in that, The specific steps for outputting the optimal target CNN network include: The images in the input simulation dataset are processed using the forward propagation algorithm to obtain the predicted stripe period; The predicted stripe period is compared with the target label in the simulation dataset, and the error between the two is calculated using a loss function. Based on the calculated error, the parameters of the initial CNN network are iteratively updated using the backpropagation algorithm combined with the gradient descent optimizer according to a preset learning rate until convergence is achieved, and the optimal target CNN network is output. In this process, a constant correction value is introduced into the denominator of the floating-point division operation performed by the gradient descent optimizer. The constant correction value is preset based on the need to prevent numerical underflow. The preset learning rate is set based on the trade-off between the convergence speed in the early stage of training and the optimization stability in the late stage of training.

6. The artificial intelligence detection method for the perpendicularity of the working surface of a beam splitter prism according to claim 1, characterized in that, The specific steps for extracting the truncated measurement image from the experimental interferogram include: Remove the edge distortion regions of the experimental interferogram and retain the effective interferogram with a size larger than the input dimension of the optimal target CNN network; By traversing the effective interference region through a sliding window, the gray-level variance of local pixels is extracted as a contrast feature, and the Euclidean distance from the center of the window to the geometric center of the global effective aperture is calculated as a positional deviation feature. A multi-dimensional region evaluation function is established by weighting and summing the spatial penalty coefficient set for the position deviation feature with the contrast feature. The intercepted measurement map is obtained based on the coordinates of the maximum value of the multi-dimensional region evaluation function.

7. The artificial intelligence detection method for the perpendicularity of the working surface of a beam splitter prism according to claim 1, characterized in that, The specific steps for deriving the fringe period formula based on the interference theory model and the correlation of interference parameters include: Based on the interference theory model established on the principle of equal thickness interference, it is determined that the optical path difference generated by the two remaining reflected beams at the meeting point satisfies the optical path difference formula. The spatial distribution of light intensity received on the target surface of the photodetector (7) after the two coherent beams are superimposed follows the interference light intensity formula. The interference light intensity formula includes the optical path difference determined according to the optical path difference formula. When the cosine term in the interference light intensity formula reaches its maximum value, constructive interference occurs, forming a bright fringe center. The corresponding physical medium thickness condition conforms to the bright fringe center formula. By analyzing the evolution of the medium thickness difference between adjacent bright fringe centers determined by the bright fringe center formula, and by establishing a mathematical relationship between the period and the minute angle using geometric trigonometric functions, the stripe period formula is derived.

8. The artificial intelligence detection method for the perpendicularity of the working surface of a beam splitter prism according to claim 1, characterized in that, The specific steps of substituting the final fringe period into the prism angle formula derived from the fringe period formula and calculating the actual angle of the working surface using the arctangent function include: By performing a mathematical transformation on the stripe period formula, the prism angle formula is derived. Substitute the final fringe period, the incident laser wavelength and refractive index corresponding to the externally input detection environment into the prism angle formula; A non-zero division check logic is added to determine whether the final stripe period approaches zero. After passing the check, the minute included angle value is calculated using the arctangent function. The actual angle of the working surface is obtained by superimposing the constant reference value of the ideal orthogonal state on the value of the small included angle, so as to complete the quantitative detection of the perpendicularity error of the working surface of the beam splitter (8) under test.

Images